PaLM 2, Google’s newest large language model (LLM), was introduced recently. It will power Google’s revamped Bard chat tool and serve as the basis model for the majority of the new AI capabilities announced today. The model was trained on a vast quantity of math and scientific materials, as well as mathematical equations, and was built on top of Google’s newest JAX and TPU v4 infrastructure. It also has enhanced support for developing and debugging code, and it was trained on 20 different programming languages, including popular ones like JavaScript and Python, but also Prolog, Verilog, and Fortran. PaLM 2 is the foundation for Codey, Google’s specialized model for coding and debugging, which is also being released today.
Google stated today that PaLM 2 was trained on a corpus of more than 100 languages, allowing it to “excel at multilingual tasks.” Google also has a family of models that can operate on smartphones, including Codey, Med-PaLM, Sec-PaLM, and a smaller PaLM 2. Google admits that they have taken a deliberate approach to releasing AI features, but that they aim to create these tools responsibly and with safety in mind. We don’t know how well it runs or how it handles edge cases because we couldn’t test it before today’s announcement.
